Perplexity's Portable Computer Local AI Agent Comes to Windows RTX PCs

Portable Computer is the on-device version of Perplexity's agentic Computer platform, and its defining decision is where each step of a task actually executes. By default, every task starts on the device, using an NVIDIA GPU-accelerated local model to plan, retrieve files, and carry out multistep work without leaving the machine ^[1]. Work completed locally draws down zero Perplexity Computer billing credits, and the system only sends an individual step to a more powerful cloud model after the user explicitly grants permission for that step ^[2].

Perplexity's VP of Engineering for Infrastructure and Enterprise, Nate Kupp, says the local app 'carries the exact same UI as the original Computer platform and includes every feature needed for local inferencing' ^[4]- the goal being to remove the historical friction of standing up a local AI stack rather than ship a stripped-down companion app. That parity extends to the agent's reach: Portable Computer adds local MCP server support so it can drive other installed desktop apps on-device, plus a new scheduler that runs recurring tasks on the PC while the user is away, with out-of-the-box connectors for Google Drive, Gmail, Slack, GitHub, Outlook, OneDrive, Word, and PowerPoint ^[3]. That connector list continues a build-out Perplexity started months earlier, when Computer first shipped inside Microsoft Teams and then gained native add-in side panels for Word, Excel, PowerPoint, and Outlook ^[5].

Reception outside the official channels echoed that same emphasis on step-level routing rather than raw model quality: independent commentary on X.com zeroed in on the local/cloud step split as the notable design choice, while early hands-on reaction on YouTube was immediately enthusiastic about trying the on-device agent - though as with any day-one launch, that reaction is still first impressions rather than settled judgment.

Perplexity's own benchmarking positions Portable Computer as the strongest fully local agent harness available today, not merely a privacy-friendly downgrade. On a 53-task Local Knowledge Work Bench, Computer running Qwen 3.8 27B scored 82.6% versus 77.6% for Pi Harness and 74.0% for Hermes, and Perplexity's own post-trained PPLX 27B model pushed that to 85.4% ^[2]. On BrowseComp, a web-research benchmark, Portable Computer hit 66.7% accuracy against Pi Harness's 50.2% and Hermes's 43.9% - while using 51% less wall time and 70% fewer tokens than Pi ^[2]. A multimodal document-understanding test showed an even wider gap: 65.1% for Computer versus 34.6% for Hermes and just 13.9% for Pi ^[2].

The trade-off shows up clearly on harder, more agentic tasks. On Terminal Bench 2.1, a coding benchmark, the fully local Qwen configuration scored 59.6% at zero marginal cost; escalating to a cloud frontier model on the harder steps raised the score to 73.0%, but at roughly $0.415 per task ^[6]. That gap is the real product of Portable Computer's local-first routing - it isn't that local models are broadly competitive with frontier cloud models, it's that Perplexity's harness lets users buy back most of the capability gap selectively, task by task, instead of paying token costs on every single step.

Enterprise security specialists reviewing Portable Computer's cloud-escalation design are not taking Perplexity's permission-prompt model at face value. Aman Mahapatra, Chief Strategy Officer at Tribeca Softtech, argues that running the control plane locally does shift the escalation decision on-device, but says the current mechanism is 'consent, not control' - it lacks egress proxies, DLP inspection, deterministic classification, and immutable audit logging that enterprises typically require before trusting an agent with sensitive data ^[7]. Justin Greis, CEO of Acceligence, frames the risk as a branding problem as much as a technical one, warning that 'local-first should not be confused with local-only' and that the distinction 'is going to matter tremendously' for organizations that assume the product never touches the cloud ^[7].

Perplexity disputes the sharpest version of that critique. Communications manager Beejoli Shah says content sitting in a local document 'can't authorize an escalation by itself, nor can it override product controls,' and that any escalation requires an explicit, single-use approval from the user ^[7]. Mike Wilkes, enterprise CISO at Aikido Security, complicates the picture further by noting that the obvious blunt fix - simply blocking the agent's network access - would defeat the product's usefulness, since he 'would not expect either the agents or their users to be particularly happy' with that trade-off ^[7]. The unresolved question is less about whether Portable Computer can be secured and more about whether a consumer-grade permission prompt is the right control surface for regulated enterprise data at all.

Portable Computer's zero-token-cost pitch obscures a real upfront bill. The minimum spec is a GPU with 24GB or more VRAM - RTX 3090, RTX 3090 Ti, RTX 4090, RTX 5090, or an RTX PRO Blackwell workstation card ^[8]- and NVIDIA's own DGX Spark, the system Portable Computer first launched on, costs roughly $4,699 ^[9]. Gartner VP analyst Nader Henein sums up the market skepticism bluntly: 'until we see the price, it's going to be hard to get excited about this' ^[7], while LexisNexis Risk Solutions Group CISO Flavio Villanustre points out that 'the hardware demands are quite steep, especially with the current hardware costs including RAM, GPU' ^[7].

That tension is already visible among people who cleared the hardware bar. Early adopters running the DGX Spark version report the app pulling in roughly 100GB of unified memory just to load a 30GB inference container and a 27GB local model - a serious commitment even on hardware built for the job - while others in the same community pushed back on the fact that work running entirely on a device they already purchased is still gated by a Perplexity subscription and credit allotment. NVIDIA is the clearest winner in that arrangement regardless of how the credit debate resolves: Portable Computer is a flagship reason to buy a high-VRAM RTX or RTX PRO card, converting Perplexity's local-AI pitch directly into GPU demand.